4,135 videos, and the finding that moved the target.
The question
Affiliate managers have one expensive question: which creators already recommend our products to a real audience, and are not in our program?
Not "find me influencers." A dozen tools filter a database by follower count. Creators who advocate unprompted convert differently from ones you recruit cold, and finding them means reading what people actually said on camera. That is a language problem wearing a filtering problem's clothes, which is why the existing tools walk past it.
I ran go-to-market at a record label before this, so it is a question I have been on the buying side of.
What the survey found
11 channels, 4,135 videos, tagged by hand and by model.
I went in with two creator states: already a partner, or an advocate you can sell to. The data showed a third. Creators who mention a product through retail, with no direct relationship at all, and that group dwarfed the one I had built the model around. 22 direct relationships against 214 through retail. At brand level, 79 appeared via retail against 16 direct.
That is not a rounding error. It is a different product. A tool built on the two-state assumption would have served the smallest group and looked correct the whole time.
Where it stands
Parked at seven commits, because the product is not there yet. Three states mean a different qualification path than the one I had scoped, and building the version I was holding would have shipped the wrong shape.
The survey is the asset. Picking this back up starts from what the data actually showed instead of what I assumed, which is a considerably better place to start than where I was.